The Compute Heat Rate - AI, Data Centers, and the Future of Power Market Pricing

Greetings from Mexico!
 In this episode, we dive deep into a fascinating new metric called the Compute Heat Rate (CHR) and explore its potentially profound implications for the future of electricity prices and the power grid
.
With the explosive growth of AI, power grids are facing unprecedented demands. For example, Texas's ERCOT grid operator recently projected that load could max out at a staggering 319,650 megawatts by 2030, driven largely by data centers
. We are already seeing the impact in markets like PJM, where data load growth has blown up capacity revenues to the tune of an estimated $23 billion in costs over the next three years
.
But what happens to the actual energy prices? That is what the CHR attempts to answer by asking: at what price would data centers elect NOT to consume power?
Introduced by industry veteran Hans Royal, the CHR measures the maximum electricity price a data center operator can rationally pay before their computing tasks become uneconomic
. Because AI creates enormous economic value, these data centers are incredibly inflexible and willing to pay massive premiums for power
. While traditional large loads like steel or aluminum producers will typically shut down when prices hit $40 to $120 per megawatt hour
, Royal estimates that AI data centers have a blended CHR of approximately 6,350permegawatthour
.Forhighlycritical,just−in−timeAIinferenceservices,theymightnotcurtailpoweruntilpriceshitover∗∗53,000 per megawatt hour**!
Watch to learn more about:
The massive gap between AI load forecasts and grid realities
.
Why regulators are demanding "Flex Mosaic" and load-shifting capabilities from data centers
.
The difference between Large Language Model (LLM) training loads and peaky inference loads
.
How the incredible power density of new tech—like the Nvidia Rubin architecture, where a fridge-sized box uses the power of 65 households—could price regular consumers out of the market
.
If these data centers refuse to curtail power at any normal wholesale price, we could see massive localized demand supply imbalances
. Watch now to understand the new metric tracking this emerging grid crisis

Transcript

Data Center Growth And ERCOT Shock

Flexibility Requirements For New Loads

Compute Heat Rate And Price Elasticity

Market Impacts And What To Watch

SPEAKER_00 0:03

Greetings from Mexico, where I suffer from a pretty severe head cold, but hey, you get what you get. I recently wrote an opinion piece for RTO Insider, the online news organization that provides coverage of competitive power markets, regulations, and federal and state climate policy about a new metric that addresses the interaction between AI data centers and power markets. It's called the compute heat rate, and it's such a fascinating concept with potentially profound implications for electricity price formation. I thought I would address it here as well. We all know about the astonishing growth of data centers, but just to set one's eyes popping, let's pull out the ERCOT forecast from this month. ERCOT's record peak demand was set in August of 2023 at 85,500 megawatts. The grid operator's latest forecast currently projects 2030 load to max out at 319,650 MW, which would drop down to 107,000 megawatts if large and medium-sized loads were not included. So mostly data load. And yes, admittedly, that number lives in stupid land. It's not in any way going to be possible to triple or nearly quadruple the Texas grid in four years. But nonetheless, big is big. Even if that grid were only to grow by a measly 50%, still a heavy lift, these new loads would affect power markets. We've already seen the impact of data loads on capacity prices in BJM, where a much more modest rate of load growth than that of Texas has blown up the capacity revenues with an estimated$23 billion of costs over the three-year 2025-2028 period. Too much demand, chasing limited supply, will do that to you. That's capacity. But what about energy? What might the impact on those prices be? Well, that's what the compute heat rate, the CHR is aimed at clarifying. And the essential question the CHR asks is, how price-sensitive or data center loads? At what price would they elect not to consume power? The question is an important one because if these loads dominate certain grids or transmission-constrained locations in these grids, they may be willing to significantly outbid all of the loads and establish the locational marginal prices. Before we get into that further, though, let's look at the concept of data center flexibility, a topic that's been raised quite frequently with respect to the grid interconnection of these new large loads. That's because they push up peak demand, requiring the construction of new generation and transmission infrastructure. In response, and to address the critical resource adequacy supply-demand issue, many regulators and politicians are now saying data loads must offer flexibility and not put pressure on the system peak if they want to be connected. In Texas, for example, Senate Bill 6 stipulates that data loads can be cut off during grid emergencies prior to the enactment of rotating blackouts. Southwest PowerPool will expedite into connections if loads can commit to flexible operations, and PGM has also discussed bring your own capacity approaches. This flexibility may be achieved in some instances by shifting compute functions to other locations, or using software to alter the compute function or deploying on-site batteries. Recently, the Electric Power Research Institute rolled out a proposed common framework called Flex Mosaic that would be used by utilities and grid operators to create approaches for such flexibility, looking at issues like notification requirements, event duration, frequency of events, depth of load adjustment, ramp behavior, and availability. All of these are very similar to what the demand response industry world has required of DR assets for decades. As I wrote in the RTO piece, the animal has to perform pretty much the same old trick, but it's far more critical when it's done by an elephant than a mouse. There are also a few distinctions that must be discussed. First, there are generally two types of AI loads, those that serve the large language training models, the LLMs, and those that serve inference, which is the on-demand use of those models, once developed, to perform tasks in our daily lives. Each type of load has a different shape. LLMs are generally far flatter, with higher load factors since they run for weeks or months on end crunching data. Inference loads, by contrast, are generally peakier, which makes sense since many inference functions generally track human activities. They're active when we're active. Each of these loads will generally exhibit different price elasticities, in other words, price responsive behavior. And it's this price-responsive behavior that the CHR attempts to quantify, and the prices that stimulate curtailment for AI data centers are likely to be much higher than the estimated$100 per megawatt hour at which Bitcoin miners typically shut down. But how much higher? Well, industry veteran Hans Royal recently published a paper introducing the compute heat rate that he defines as the maximum electricity price a data center operator can rationally pay before the computation running on that electricity becomes uneconomic. He modeled the concept after the generating plant heat rate, the metric for a power plant's thermal efficiency in converting the potential heat in a given fuel, usually gas, into electricity. So the BTU is needed to generate a kilowatt hour of electric energy. Other large loads, aluminum smelters, steel producers, chemical manufacturers, they're pretty price sensitive and they often curtail at costs of between$40 and$120 per megawatt hour, a self-correcting price mechanism in the market. But AI data centers, they're different because of the potential enormous economic value they create per megawatt hour of power used. So the more valuable the compute task at hand, the greater the willingness to pay a higher price. In other words, that load becomes less elastic. And that number increases as ships and systems become more efficient in terms of compute per unit of energy. Well, here's what Royal recently found: a blended CHR of approximately$6,350 per megawatt hour. This implies, Royal surmises, that AI data center demand will not curtail at electricity prices below roughly 127 times the current wholesale average. Different types of AI workloads will exhibit different price sensitivities. Initial estimates have them ranging from about$500 per megawatt hour for training of a frontier large language training model to over$53,000 per megawatt hour for frontier inference services. The LLMs operate over longer periods of time and prices are averaged. Inference is just in time delivery of high value information for potential agentic services like trading, logistical planning, or robotic activity, and it's less flexible. Users need that information now. So naturally, supporting input power prices would be higher. Rowe concludes that because of these high opportunity costs, these massive loads, quote, will not curtail at any price level observed in the U.S. wholesale markets, unquote. Of course, these elasticities don't really matter when there's sufficient supply, but as data loads grow, the demand supply imbalance increases. At that point, the CHR dynamic does apply, especially in localized and transmission constrained areas. And Royal postulates such prices will occur once critical penetration thresholds are reached, with a CHR effect emerging suddenly once sufficient data center infrastructure reaches critical mass at specific grid nodes. If this dynamic actually does occur, it has significant implications for grid operators, utility planners, and regulators. And it certainly impacts other electricity consumers who may find themselves priced out in various markets. If it drives up day ahead or real-time prices, it will also impact forward curves. Although interesting, a just released International Energy Agency report on Energy and AI notes that recent ERCOT forward curves do not anticipate any of this dynamic yet. The CHR metric may require additional work and it's likely to continue to be refined. That's beyond my background and ability to comment on. But I think it has important implications for where we might be headed. As the EIA report comments, to put this in perspective, with the announced NVIDIA Rubin architecture, a box the size of a household refrigerator would have a peak power draw equivalent to that of around 65 households. The scale with which the AI industry plans to power these things portends massive new demand, something we've simply never seen before. This CHR metric may offer great value as a tool for tracking market costs by location over time as the use of AI accelerates and the associated technologies continue to evolve. For that alone, the concept is worth watching. Well, thanks for watching, and we'll see you again soon.